How do I design the local development environment to simulate realistic partial failures, so I can see durable execution recovering state?

Introduce an in-process fault injector that raises exceptions at specific step boundaries or causes transient I/O errors. Run the agent with a local state store and observe how Catalyst replays from the last checkpoint after each failure. This approach verifies your step idempotency under controlled conditions. Note that this simulates crashes but cannot replicate all production failure modes like network partitions or node outages, which may affect external service interactions.